The image captures a segment of an academic poster presentation featuring research by the National University of Singapore (NUS) and the CVPR community. The poster, titled "DeepCache: A novel training-free paradigm that accelerates diffusion models," is part of an exhibition focused on advancements in diffusion models. It highlights a method that dynamically compresses model architecture, improving efficiency in the sequential denoising process. The visual content includes diagrams, graphs, sample images demonstrating the technology's effects, and a detailed explanation of the underlying mechanisms. A person's hand is seen pointing at the diagrams, suggesting an in-depth discussion or explanation of the research work. The setting appears to be a conference or symposium, with other research posters visible in the background.
Text transcribed from the image:
CVPR MANUS
JUNE 17-21, 2024
National University
of Singapore
DeepCach
DeepCache:
SEATTLE WA
Introduction
A novel training-free paradigm that accelerates diffusion models
from the perspective of model architecture by dynamically compressing
diffusion models during runtime
Motivation:
Temporal redundancy observed in the sequential denoising steps of
diffusion models.
Method:
Reuse the high-level features while updating the low-level features in a
very cheap way
pay
bar with a heat
the prope
Apple d
(a) Examples of Feature Maps
(b) HeatMap for Similarity
Stable Diffu
Stable Diffusi
110
kay
blayd
LDM+G
DDPM/Bedroom
SD v1.5
(c) Ratio of steps that similarity larger than 0.95
100%
92%
80%-
70%
60%-
50%
40%-
30%
20%-
Original
Step20
10%
Step19
Step
Step
142
D₁₂
Skip Branch 1
20%
40%
60%
80%
Denoising Progress
Xt
Xt+1
100%
A person in a
helmet is riding a
skateboard
There are three
vases made of
clay on a table
A bicy
standing
a bed in
U₁
D₁
Stable
Diffusion
v1.5
D₂
U2
Skip Branch 3
U3
M
-Chrzan Path
Excited Path
M+2
BK-SDM
-Tiny
Time: 4.3395
Ours
Time: 2.571s
Time: 4.356s
Time: